Download FORUM Challenges and Limitations of Gene Expression Profiling in

60, 6 –10 (2001)
Copyright © 2001 by the Society of Toxicology
TOXICOLOGICAL SCIENCES
FORUM
Challenges and Limitations of Gene Expression Profiling
in Mechanistic and Predictive Toxicology
Mark R. Fielden and Tim R. Zacharewski 1
Department of Biochemistry and Molecular Biology, National Food Safety and Toxicology Center, Institute of Environmental Toxicology,
Michigan State University, 223 Biochemistry Building, Wilson Road, East Lansing, Michigan 48824
Received August 16, 2000; accepted October 23, 2000
drug development and aid risk assessment. Recent experiments
applied to cancer genetics have demonstrated the potential of
gene expression profiling to accurately classify disease phenotypes (Alizadeh et al., 2000; Bittner et al., 2000), thus lending
hope that expression profiling may classify and thus predict
phenotypes of toxicity. Despite these expectations, it is still
uncertain how gene expression profiling experiments will ultimately contribute to our understanding of toxicity and allow
us to realize the full potential of this new technology. Although
there has been much review and hyperbole surrounding the
potential applications of toxicogenomics, these novel and unverified approaches to toxicological problems require an
awareness of the constraints of the methodology in order to
design and interpret gene expression profiling data. Pennie et
al. (2000) have also discussed the possibilities and caveats of
gene expression profiling in the context of mechanistic and
predictive toxicology and have addressed the certainty, biological relevance, and need for validation of microarray data. The
purpose of this paper is to illustrate the current constraints of
gene expression profiling in mechanistic and predictive toxicology and to stress how current experimental designs may
confound accurate interpretation of genome-scale data. The
limitations described are not intended to discourage the application of gene expression profiling technologies to mechanistic
or predictive toxicology, but rather guide experiments that will
produce more interpretable and useful data.
RNA and protein expression profiling technologies have revolutionized how toxicologists can study the molecular basis of
adverse effects of chemicals and drugs. It is expected that these
new technologies will afford efficient and high-throughput means
to delineate mechanisms of action and predict toxicity of unknown
agents. To reach these goals, a more thorough understanding of
the constraints of the methodology is needed to design genomescale studies and to interpret the vast amount of data collected.
This paper addresses some of the limitations and uncertainties of
gene expression profiling in mechanistic and predictive toxicology
with respect to the expectations of toxicogenomics. The challenges
associated with interpreting information from large-scale gene
expression experiments in vivo is also discussed.
Key Words: gene expression profiling; mechanistic and predictive toxicology; toxicogenomics; genome analysis.
cDNA and oligonucleotide arrays and high-throughput 2-D
electrophoresis systems have quickly emerged as the premier
tools to enable genomewide analysis of gene expression at the
RNA and protein level. These new technologies are heavily
influencing drug discovery and preclinical safety in the biotechnology and pharmaceutical industry (Freeman, 2000). Toxicologists are also promoting genomic expression technologies
as a superior alternative to traditional rodent bioassays to
identify and assess the safety of chemicals and drug candidates
for human safety (Afshari et al., 1999; Nuwaysir et al., 1999;
Pennie et al., 2000). It is expected that gene expression profiling will identify mechanisms of action that underlie the
potential toxicity of chemicals and drug candidates. Other
touted applications include the identification of biomarkers of
toxicity to predict potential hazardous substances and therapeutics. Ultimately, toxicogenomics (the integration of genomics, bioinformatics, and toxicology) is expected to accelerate
Gene Expression Profiling in Mechanistic Toxicology—
A Hypothesis-generating Tool
There is a certain degree of faith that gene expression
profiling will reveal the mechanisms of action of chemicals and
drugs despite the inherent limitation of genomic and proteomic
experiments, which measure single end points (i.e., RNA or
protein levels), albeit for thousands of genes at a time. Consider the many experiments and end points that have been
employed to explain the mechanism of action of some previ-
1
To whom correspondence should be addressed. Fax: (517) 353-9334.
E-mail: [email protected].
6
FORUM
ously characterized chemicals and drugs. It is understood that
mechanisms of action are far more complex and affect more
than simply the levels of cellular macromolecules. Many toxicants affect enzyme activity, DNA integrity, redox status,
membrane integrity, and other processes that are not amenable
as yet to genomewide measurements. Although alterations in
the above processes are likely to indirectly affect the expression of genes and proteins, the question remains how we
extrapolate a mechanism of action from the one end point.
Similarly, predictive toxicology attempts to infer the potential
mechanism(s) of action of an unknown agent on the basis of
correlation to large databases of activity or expression profiles
(Hughes et al., 2000; Paull et al., 1989; Scherf et al., 2000).
Can mechanisms of action be determined or predicted from
gene expression profiling? To answer this question, we need to
first define what is meant by a mechanism of action, as the term
is often used with many connotations. The mechanism of
action of a chemical or drug is described by the series of
molecular events following interaction of a chemical with its
cellular target(s) and the subsequent alteration(s) in target
function that precedes a cascade of cellular events that ultimately leads to the observed effect. The challenge of trying to
determine the mechanism of action from measuring steadystate mRNA or protein levels is that many toxicants and drugs
initiate toxicity by binding to proteins and/or altering macromolecules (although with exceptions, as noted below), and not
by directly inducing gene expression or altering gene product
stability or turnover. For example, the mechanism of action of
acetaminophen (APAP)-induced hepatocellular necrosis is due
to cytochrome P450 – catalyzed activation of APAP to the
electrophilic NAPQI intermediate, leading to arylation and
thiol oxidation of cellular proteins. These events in turn lead to
nonspecific and/or undefined alterations in protein function and
subsequent changes in nuclear and organelle structure and
function leading to irreversible cell injury and oncotic necrosis
(Cohen and Khairallah, 1997). The mechanism of action of
APAP has been delineated through many detailed chemical and
biochemical experiments that could not have been revealed
through observation of gene expression changes alone.
This is a limited view of the complete spectrum of toxic
effects initiated by APAP, as it has been observed to cause
chromosomal aberrations, apoptotic DNA fragmentation, unscheduled DNA synthesis, oxidative stress, altered calcium
homeostasis, and inhibition of cell proliferation (Boulares et
al., 1999, and references therein). The fact that multiple cellular signaling pathways may converge to alter the expression
of the same gene products also makes it difficult to identify the
affected pathway from observing gene expression changes.
The above arguments illustrate the point that most chemicals
and drugs will act through multiple mechanisms of action that
will depend on dose, timing and duration of exposure, and cell
phenotype. Each individual mechanism represents an initiating
event which by itself is inadequate to drive progression of
toxicity, but these mechanisms together act in concert to cause
7
cell injury and/or death. Although gene expression would be
expected to be altered as a result of APAP exposure, the
changes in gene expression will reflect secondary outcomes
due to primary upstream events starting with the interaction of
APAP with its target protein(s). Therefore, our ability to define
the mechanism of action of a compound using gene expression
profiling technologies will be highly limited in resolution. In
the best-case scenario, gene expression changes in cellular
perturbation experiments will lead to many new testable hypotheses that will require subsequent molecular and biochemical experiments to reveal and confirm precise mechanisms of
action.
Ultimately, being able to define the mechanisms of organismal toxicity will depend on our understanding of cellular and
tissue level effects and how they are related to the molecular
changes in target cells. Because changes in gene expression do
not necessarily imply toxicity, gene expression profiling experiments need to be integrated into larger studies that examine
multiple end points at the molecular, cellular, tissue, and physiological levels in the context of the whole organism. As noted
(Pennie et al., 2000), this creates a further challenge in trying
to integrate knowledge at all levels of biological organization,
and highlights the need for an interdisciplinary approach in
mechanistic toxicology.
In some instances of toxicity, a direct and primary response
affecting gene expression and subsequent initiation of toxicity,
due likely to a receptor-mediated pathway, may be used to
explain the mechanism of action of chemicals, including nongenotoxic carcinogens or endocrine disrupters. This will be
particularly true for therapeutics and drug candidates, as it has
been estimated that close to 50% of marketed drugs act through
receptors (Drews, 2000). These observations will be complicated, of course, by parallel mechanisms of toxicity that may or
may not be receptor mediated, yet may augment the receptormediated events. The challenge will be to distinguish the
therapeutic affects from the pathological changes. This will
require establishing time-dependent relationships between
dose and toxicity, which may or may not be linear. Where
alterations in gene expression precede or coincide with toxicity, our ability to understand the mechanism of action will be
limited to our understanding of the pathways that regulate
transcription of the affected genes and their kinetics of expression. This is currently a major limitation in understanding why
a particular gene or cluster of genes is observed to be up- or
down-regulated, since only a small fraction of the estimated
100,000⫹ human genes have been studied at the level of
transcriptional regulation. Combining the identification of gene
regulatory elements with expression profiles in microarray
experiments (Barzma et al., 1998; Zhang, 1999) represents an
industrious approach to begin to understand what transcription
factors and upstream signaling molecules are governing the
observed response in gene expression following chemical or
drug exposure.
Gene expression profiling has possibly a greater potential to
8
FORUM
reveal modes of action through the analysis of secondary
responses and/or the series of contingent regulatory events
induced by chemical or drug exposure. The mode of action of
a chemical or drug can be described, in part, by a fundamental
obligatory step directing toxicity, or adverse cell fate, be it
reversible cell injury, apoptotic or oncotic necrosis, or malignant transformation. Farr and Dunn (1999) have noted that
organismal manifestations of toxicity can be explained by
combinations of a limited number of cellular outcomes from a
limited number of cell and/or tissue types. Furthermore, multiple mechanisms of action may converge at common points to
trigger the same molecular response. If this is true, then the
number of possible modes of action will be limited to the
number of molecular responses that can drive the obligatory
step toward a discrete cellular outcome. It follows, then, that
gene expression profiles cannot be used as an explanation or
predictor of toxicity unless correlated with an adverse effect.
Again, this underscores the need to integrate genomic experiments with experiments examining effects at higher levels of
biological organization that are intended to assess toxicity in
the context of the whole organism. By understanding the gene
expression changes that direct a unique cellular outcome (i.e.,
the mode of action), we can begin to use gene expression
profiles to explain and potentially predict toxicity.
Predictive Toxicology—Fact or Fiction
It has been proposed that each chemical that acts through a
particular mechanism of action will induce a unique and diagnostic gene expression profile under a given set of conditions
(Nuwaysir et al., 1999). Indeed, proof-of-principle experiments
in S. cerevisiae have revealed that the response to inhibitory
compounds mimics the loss of function of its target or pathway
for at least six compounds (Hughes et al., 2000, and references
within; Marton et al., 1998). For example, genetic disruption of
calcineurin in S. cerevisiae resulted in a gene expression profile
highly correlated with the expression profile of wild-type cells
treated with FK506 or cyclosporin, antagonists of the calcineurin-signaling pathway. To estimate the significance of the
relationship, the FK506 treatment profile was compared to
more than 40 randomly selected deletion strains or drug-treated
cells and found to be uncorrelated (Marton et al., 1998).
Whether predictive patterns in gene expression can be observed in mammalian systems remains to be shown, although
preliminary studies suggest they can (Blanchard et al., 2000;
McMillian et al., 2000). Therefore, there is significant potential
for chemicals and drugs to be classified based on the similarity
of their induced gene expression profile by comparison with
expression profiles induced by chemicals or drugs with known
mechanism of action using multivariate statistical methods and
correlation metrics. In some cases, however, their classification
may be limited to the affected signaling or metabolic pathway
rather than by target protein in the pathway. By extension of
this observation, gene expression profiles are anticipated to
produce knowledge of a subset of commonly regulated genes
that can be used as biomarkers to predict modes of action.
While it has been pointed out that the number of possible
patterns of differential gene expression, even when expressed
as binary variables, is enormous (Farr and Dunn, 1999), subtle
differences in the number and magnitude of gene expression
changes have proven to be sufficient to classify expression
profiles into distinct clusters when applied to S. cerevisiae
(Hughes et al., 2000). The utility of this approach, however,
may be lost when outside the context of a large database, or
compendium, of expression profiles, as subtle changes in relative expression level (i.e., less than 2-fold) are usually considered unreliable in isolation (Hughes et al., 2000). Based on
gene expression profiles of yeast mutants, it has been estimated
that there exist 300 to 700 distinct full genome transcriptional
patterns from a full set of 5000 yeast deletion mutants profiled
under a single condition (Hughes et al., 2000). Although this
was a crude prediction, an extrapolation to mammalian systems
may predict substantially more distinct transcriptional patterns
under a single condition. Classifying transcriptional responses
into distinct diagnostic clusters may prove more problematic if
responses under different conditions do not extrapolate under
different conditions. For example, transcriptional responses
may differ between one target cell and another, from cell
culture to in vivo conditions, or from rodent models to humans.
Thus, the predictive power of gene expression profiling may be
limited to the model system employed and the prototypical
compound with known mechanisms used to generate the diagnostic expression profile. As yet there is no published data to
support that predictive expression profiles will extrapolate to
other tissues or in vivo settings.
The Challenge of Interpreting Gene Expression Data
Currently, there is a significant knowledge gap in our understanding of the molecular events that govern toxicologically
relevant outcomes. In any event, the changes in gene expression directing cell fate will reflect, in part, an active physiological response that is nontoxic. These responses may include,
but are not limited to, host– defense responses (e.g., acute
phase proteins, cytokines, DNA repair enzymes), adaptive responses (e.g., hyperplasia, metaplasia, hypertrophy, atrophy),
and regenerative or protective responses (proliferation, differentiation). In addition, there will be secondary responses following toxicity that will reflect pathology as a result of disturbances in cell function. These responses are likely to be
idiosyncratic and diverse across cell types due to the interaction of pathological responses with the physiological mechanisms of detoxification and repair that are cell specific. Again,
the challenge then lies in differentiating the physiological
responses from the diagnostic pathological changes in light of
confounding experimental artifacts inherent in the model system and the experimental design.
Consider, for example, an experiment designed to measure
FORUM
time- or dose-dependent changes in gene expression following
an EC 50 dose of a cytotoxic chemical in cultured cells. When
administering a dose that kills half the cell population, the
measured response (i.e., mRNA or protein level) in the affected
culture will be a combination of multiple factors, including the
gene expression changes in dying cells due to treatment, adaptive changes in surviving cells due to treatment, and normal
responses in living cells due to adjacent necrotic cells. This
would be particularly relevant in vivo, as necrosis can induce a
regenerative or inflammatory response in some populations of
unaffected or resistant cell types. The heterogeneous responses
are likely to be highly dependent on the tissue or cell type
affected, again highlighting the limitation of extrapolating one
model system to another. Measuring gene expression changes
following sublethal exposure concentrations may be more
likely to reveal treatment-induced changes that initiate toxicity
before heterotypic cellular responses obscure interpretation.
This will require a complete characterization of the full doseand time-response relationship, including a qualitative description of cellular changes as correlates.
Artifactual complications may also apply to other classes of
chemicals, particularly chemicals that act through receptormediated pathways, as receptor expression is usually restricted
to discrete cell types. Subsequent changes in paracrine signaling may have dramatic effects that could lead to misinterpretation of gene expression profiles in cultured cells. This would
also be particularly relevant in vivo, where cellular complexity
plays a dominant role in adaptation and defense, or when target
tissues are affected secondary to primary targeting of a proximal endocrine gland such as the pituitary or thyroid. Furthermore, when analyzing gene expression profiles from whole
tissues as part of a whole-animal toxicology study, the relevant
gene expression changes in the specific cell types targeted by
chemical or drug may be masked or diluted by the benign
changes in surrounding cell types. For example, consider the
cell type–specific toxicity of alloxan or streptozotocin on the ␤
cells of the pancreas and the fact that the ␤ cells represent less
than 2 % of the pancreatic cell population. The ability to detect
changes in gene expression within 2 % of an RNA sample
derived from whole pancreas is likely below the limits of
sensitivity of current genomic profiling platforms. Compensatory changes in other, more abundant cell types may also
negate any changes in the targeted cell and could even result in
the opposite conclusion regarding message or protein abundance. Being able to measure gene expression profiles in
individual targeted cells or cell types, by using laser capture
microdissection, for example, would be more desirable in these
instances (Emmert-Buck et al., 2000; Luo et al., 1999). However, prior knowledge of the target tissues and/or cell types
from pathology studies are typically required for this level of
investigation. This would preclude its utility in higher throughput predictive assays that are currently desired, but would
prove useful for mechanistic studies. Reducing the number of
observations (i.e., gene expression profiles) and correlating
9
them with a binary response (i.e., apoptosis, DNA damage)
may allow for the identification of a more robust set of predictive markers with utility in higher throughput systems.
Realization of such a scenario will be heavily dependent on the
standards of known modes of action that are available, the
reproducibility of the model system, statistical robustness of
the data, and the application of multivariate methods of analysis to reduce the data set into a comprehensible and manageable number of components for purposes of classification.
Other limitations to consider arise when adverse cellular or
tissue functions are observed in the absence of discrete cellular
outcomes such as cell injury or death. In these instances, subtle
changes in cell function, such as reduced responsiveness to
endocrine signals or altered secretion or production of signaling molecules, will be more difficult to observe, as relevant
changes may be transient, posttranslational, and/or in non–
target organs. Many endocrine disruptors will likely fall into
this category. Perhaps the greatest source of complexity and
variability in gene expression profiling experiments in vivo
stems from non–treatment-related phenomena, or intrinsic
variability, which is difficult, if not impossible, to control and
reproduce. Normal fluctuations in gene expression will occur
as a result of differences in age, gender, temperature, light,
diet, and hormonal status. Although age, gender, and the external environment can be tightly controlled within experiments, comparisons between laboratories using similar treatment protocols may be more challenging when environmental
factors are not strictly adhered to. Differences in nutritional or
hydration status, time of last meal, hormonal fluctuations during estrus, and seasonal and light-induced changes in hormone
levels are more difficult to control within experiments. Such
intrinsic variation is likely to interact with timing, duration, and
frequency of treatments to alter the observed response in gene
expression. As with any experiment designed to test a hypothesis, there must be sufficient replication to assure certainty in
the experimental results.
The expectation that toxicogenomics will enable us to define
mechanisms of action and predict toxicity of unknown agents
is supported by recent studies in lower eukaryotes. However,
our current ability to define a mechanism of action or accurately predict toxicity in mammalian systems is still in its
infancy. Incorporating genomic experiments into larger studies
designed to assess effects at higher levels of biological organization is a must if one is to begin to understand and predict
organismal outcomes and possibly incorporate gene expression
data into mechanism-based risk assessment. The progression of
expression profiling into whole-animal studies also presents a
higher level of complexity that challenges our understanding of
biological systems and the interpretation of what changes in
gene expression are relevant. It is expected that experience and
interdisciplinary collaborations will continue to advance the
utility of gene expression profiling in mechanistic and predictive toxicology. However, continued discussion, debate, and
10
FORUM
the sharing of knowledge and data are vital for toxicogenomics
to move ahead rapidly.
REFERENCES
Afshari, C. A., Nuwaysir, E. F., and Barrett, J. C. (1999). Application of
complementary DNA microarray technology to carcinogen identification,
toxicology, and drug safety evaluation. Cancer Res. 59, 4759 – 4760.
Alizadeh, A. A., Eisen, M. B., Davis, R. E., Ma, C., Lossos, I. S., Rosenwald, A.,
Boldrick, J. C., Sabet, H., Tran, T., Yu, X., Powell, J. I., Yang, L., Marti, G. E.,
Moore, T., Hudson, J., Jr., Lu, L., Lewis, D. B., Tibshirani, R., Sherlock, G.,
Chan, W. C., Greiner, T. C., Weisenburger, D. D., Armitage, J. O., Warnke, R.,
Levy, R., Wilson, W., Grever, M., Byrd, C., Botstein, D., Brown, P., and Staudt,
L. M. (2000). Distinct types of diffuse large B-cell lymphoma identified by gene
expression profiles. Nature 403, 503–511.
Barzma, A., Jonassen, I., Vilo, J., and Ukkonen, E. (1998). Predicting gene
regulatory elements in silico on a genomic scale. Genome Res. 8, 1202–1215.
Blanchard, K., DiSorbo, O., Burris, R., Dunn, R., Farr, S., and Stoll, R. (2000).
Toxicogenomics: Understanding the use of microarrays for toxicology studies in vivo. Toxicologist 54, 195.
Bittner, M., Meltzer, P., Chen, Y., Jiang, Y., Seftor, E., Hendrix, M., Radmacher,
M., Simon, R., Yakhini, Z., Ben-Dor, A., Sampas, N., Dougherty, E., Wang, E.,
Marincola, F., Gooden, C., Lueders, J., Glatfelter, A., Pollock, P., Carpten, J.,
Gillanders, E., Leja, D., Dietrich, K., Beaudry, C., Berens, M., Alberts, D., and
Sondak, V. (2000). Molecular classification of cutaneous malignant melanoma
by gene expression profiling. Nature 406, 536 –540.
Boulares, H. A., Giardina, C., Navarro, C. L., Khairallah, E. A., and Cohen,
S. D. (1999). Modulation of serum growth factor signal transduction in Hepa
1– 6 cells by acetaminophen: An inhibition of c-myc expression, NF-kappaB
activation, and Raf-1 kinase activity. Toxicol. Sci. 48, 264 –274.
Cohen, S. D., and Khairallah, E. A. (1997). Selective protein arylation and
acetaminophen-induced hepatotoxicity. Drug Metab. Rev. 29, 59 –77.
Drews, J. (2000). Drug discovery: A historical perspective. Science 287,
1960 –1964.
Emmert-Buck, M. R., Gillespie, J. W., Paweletz, C. P., Ornstein, D. K., Basrur,
V., Appella, E., Wang, Q. H., Huang, J., Hu, N., Taylor, P., and Petricoin,
E. F., III (2000). An approach to proteomic analysis of human tumors. Mol.
Carcinog. 27, 158 –165.
Farr, S., and Dunn, R. T., II (1999). Concise review: Gene expression applied
to toxicology. Toxicol. Sci. 50, 1–9.
Freeman, T. (2000). High throughput gene expression screening: its emerging
role in drug discovery. Med. Res. Rev. 20, 197–202.
Hughes, T. R., Marton, M. J., Jones, A. R., Roberts, C. J., Stoughton, R.,
Armour, C. D., Bennett, H. A., Coffey, E., Dai, H., He, Y. D., Kidd, M. J.,
King, A. M., Meyer, M. R., Slade, D., Lum, P. Y., Stepaniants, S. B.,
Shoemaker, D. D., Gachotte, D., Chakraburtty, K., Simon, J., Bard, M., and
Friend, S. H. (2000). Functional discovery via a compendium of expression
profiles. Cell 102, 109 –126.
Luo, L., Salunga, R. C., Guo, H., Bittner, A., Joy, K. C., Galindo, J. E., Xiao,
H., Rogers, K. E., Wan, J. S., Jackson, M. R., and Erlander, M. G. (1999).
Gene expression profiles of laser-captured adjacent neuronal subtypes. Nat.
Med. 5, 117–122.
Marton, M. J., DeRisi, J. L., Bennett, H. A., Iyer, V. R., Meyer, M. R., Roberts,
C. J., Stoughton, R., Burchard, J., Slade, D., Dai, H., Bassett, D. E., Jr.,
Hartwell, L. H., Brown, P. O., and Friend, S. H. (1998). Drug target
validation and identification of secondary drug target effects using DNA
microarrays. Nat. Med. 4, 1293–1301.
McMillian, M., Ciervo, J., Li, L., Dunn, R., Fairfield, E., Farr, S., and Johnson,
M. (2000). Comparison of toxicant-induced gene expression patterns in
HepG2 human hepatoma cells. Toxicologist 54, 384.
Nuwaysir, E. F., Bittner, M., Trent, J., Barrett, J. C., and Afshari, C. A. (1999).
Microarrays and toxicology: The advent of toxicogenomics. Mol. Carcinog.
24, 153–159.
Pennie, W. D., Tugwood, J. D., Oliver, G. J., and Kimber, I. (2000). The
principles and practice of toxicogenomics: Applications and opportunities.
Toxicol. Sci. 54, 277–283.
Paull, K. D., Shoemaker, R. H., Hodes, L., Monks, A., Scudiero, D. A.,
Rubinstein, L., Plowman, J., and Boyd, M. R. (1989). Display and analysis
of patterns of differential activity of drugs against human tumor cell lines:
Development of mean graph and COMPARE algorithm. J. Natl. Cancer
Inst. 81, 1088 –1092.
Scherf, U., Ross, D. T., Waltham, M., Smith, L. H., Lee, J. K., Tanabe, L.,
Kohn, K. W., Reinhold, W. C., Myers, T. G., Andrews, D. T., Scudiero,
D. A., Eisen, M. B., Sausville, E. A., Pommier, Y., Botstein, D., Brown,
P. O., and Weinstein, J. N. (2000). A gene expression database for the
molecular pharmacology of cancer. Nat. Genet. 24, 236 –244.
Zhang, M. Q. (1999). Large-scale gene expression data analysis: A new
challenge to computational biologists. Genome. Res. 9, 681– 688.